feat(experiments): controlled baseline-vs-candidate eval harness for …

[Matthew (theworker02)]

…#132

Adds experiments/skillopt-eval: a pluggable-runner experiment that splits scenarios into seeded train/validation/test sets, gates bounded skill edits behind strict held-out validation improvement (fail-closed), and emits paired McNemar + bootstrap stats with full provenance as raw_results.csv / results.json / RESULTS.md plus raw trajectories.

Runners: 'mock' (deterministic offline dry-run, no keys/network) and 'superpowers' (wraps the existing isolated-overlay adapter at a pinned SHA; baseline and candidate share harness/model/settings by construction). Deterministic offline tests cover gate accept/reject, fail-closed errors, learning-rate bounding, source non-mutation, artifacts, and A/A no-claim.

[Yifan Yang (Yif-Yang)]

Reviewed head ce4e6f7c5629e4a15bb4032ab0f7064151a2fd81 against current main. This provides useful candidate-selection orchestration around the existing adapter and evalkit, but I am holding merge for these production-path gaps:

• The live runner gives baseline, candidate and successive attempts the same scenario/seed/pin trajectory filename. Candidate output overwrites baseline evidence, and the stored condition/split are empty. Baseline validation records are also absent from raw_results.csv. Please preserve a separately labeled artifact for every invocation and retain the reference used for each gate decision.
• With optimize_steps=0, two runs of the same unchanged baseline can produce optimized=false yet quality_improved=true. An offline provider-boundary reproduction with one discordant pair returned exact McNemar p=1 but still emitted the candidate-improvement verdict. The report needs to preserve the declared A/A no-claim invariant and distinguish observed run variation from a candidate effect.
• A validation batch with one pass and one timeout is accepted over a zero baseline. The timeout is counted as a failed row, but this differs from the candidate-level guarantee in test_validation_fails_closed_on_runner_error. Please align the implementation, coverage and documented fail-closed contract.

The existing focused tests pass: 150 tests plus 84 subtests. The independent reproductions exercised the actual Superpowers adapter, local pinned synthetic checkout, judges and experiment loop; only the external Claude process was substituted. No paid model calls or model-efficiency claims are involved. These require author revisions, so this is not a merge recommendation.